Getting a Useful Trial Sample

A vendor's marketing description of a dataset — "tick-level global equity trades, 20 years of history" — tells you almost nothing about whether it's actually usable for the specific question you're researching. Before committing to a full license or a lengthy cleaning pipeline, it's worth requesting a small trial sample and deliberately choosing what it should cover, rather than accepting whatever slice the vendor defaults to.

A useful trial sample is chosen to stress-test the exact edge cases that matter for the intended use, not just to "look at the data." That means including a known corporate action (a split or a merger) to see if the data handles it correctly, a period spanning a known market-stress event (to check for gaps or bad prints exactly when they'd matter most), and, if the study depends on smaller or less liquid names, a few names near the bottom of the universe rather than only the largest, most reliably-clean tickers. A trial restricted to five years of clean large-cap data will pass every quality check and still fail badly once the full order is delivered and turns out to have ragged coverage for the small-cap names the actual strategy needs.

The trial is also the cheapest point to check practical details that are expensive to discover later: point-in-time-ness (does the data reflect what was knowable on that date, or is it silently restated with hindsight), timestamp granularity, and how the vendor represents missing or halted data, since a placeholder value silently mixed in with real values can corrupt a backtest without ever throwing an error.